AI chip race: Google says its Tensor chips compute faster than Nvidia's A100

It also says that it has a healthy pipeline for chips in the future.
Ameya Paleja
Google's Bay View campus in California
Google's Bay View campus in California


Search engine giant Google has claimed that the supercomputers it uses to develop its artificial intelligence (AI) models are faster and more energy efficient than Nvidia Corporation's. While processing power for most companies delving into the AI space comes from Nvidia's chips, Google uses a custom chip called Tensor Processing Unit (TPU).

Google announced its Tensor chips during the peak of the COVID-19 pandemic when businesses from electronics to automotive faced the pinch of chip shortage. While the chip was initially supposed to power its Pixel smartphone, the company has undoubtedly made rapid strides in its development. It has been using them to power its AI research.

AI-designed chips to further AI development

Interesting Engineering reported in 2021 that Google used AI to design its TPUs. Google claimed that the design process was completed in just six hours using AI compared to the months humans spend designing chips.

For most things associated with AI these days, product iterations occur rapidly, and the TPU is currently in its fourth generation. As Microsoft stitched together chips to power OpenAI's research requirement, Google also put together 4,000 TPUs to make its supercomputer.

AI chip race: Google says its Tensor chips compute faster than Nvidia's A100
Stock image of a super computer with thousands of chips

Since these supercomputers are assigned massive tasks that a single chip cannot complete, the thousands of processing centers need to communicate among themselves. In a recent scientific paper, Google claimed it had custom-developed optical switches.

Google's PaLM model, the largest model it has publicly spoken about, was trained using a supercomputer consisting of 4,000 connected TPUs and a processing time of over 50 days. Google claims that its supercomputer can reconfigure connections between the chips on the fly, allowing it to extract performance gains from the assembly.

Google also claimed in the paper that its TPUs are 1.7 times faster than the A100 chips from Nvidia, which power most AI applications. The chips are also 1.9 times more energy efficient than the A100, making Google's AI processing greener.

While Microsoft brags about powering OpenAI's research using its supercomputer services, Google has Midjourney, among its famous customers, that has made advanced AI tools utilizing the processing power.

As the race for improved AI heats up, Nvidia has revealed its H100 chip with computing prowess. Google hasn't commented on how its fourth-generation TPU fares against the H100 but has said that it has a healthy pipeline of future chips to take on the competition.

Add Interesting Engineering to your Google News feed.
Add Interesting Engineering to your Google News feed.
message circleSHOW COMMENT (1)chevron
Job Board